unregister_key_type() has code to mark a key as dead and make it unavailable in
one loop and then destroy all those unavailable key payloads in the next loop.
However, the loop to mark keys dead renders the key undetectable to the second
loop by changing the key type pointer also.
Fix this by the following means:
(1) The key code has two garbage collectors: one deletes unreferenced keys and
the other alters keyrings to delete links to old dead, revoked and expired
keys. They can end up holding each other up as both want to scan the key
serial tree under spinlock. Combine these into a single routine.
(2) Move the dead key marking, dead link removal and dead key removal into the
garbage collector as a three phase process running over the three cycles
of the normal garbage collection procedure. This is tracked by the
KEY_GC_REAPING_DEAD_1, _2 and _3 state flags.
unregister_key_type() then just unlinks the key type from the list, wakes
up the garbage collector and waits for the third phase to complete.
(3) Downgrade the key types sem in unregister_key_type() once it has deleted
the key type from the list so that it doesn't block the keyctl() syscall.
(4) Dead keys that cannot be simply removed in the third phase have their
payloads destroyed with the key's semaphore write-locked to prevent
interference by the keyctl() syscall. There should be no in-kernel users
of dead keys of that type by the point of unregistration, though keyctl()
may be holding a reference.
(5) Only perform timer recalculation in the GC if the timer actually expired.
If it didn't, we'll get another cycle when it goes off - and if the key
that actually triggered it has been removed, it's not a problem.
(6) Only garbage collect link if the timer expired or if we're doing dead key
clean up phase 2.
(7) As only key_garbage_collector() is permitted to use rb_erase() on the key
serial tree, it doesn't need to revalidate its cursor after dropping the
spinlock as the node the cursor points to must still exist in the tree.
(8) Drop the spinlock in the GC if there is contention on it or if we need to
reschedule. After dealing with that, get the spinlock again and resume
scanning.
This has been tested in the following ways:
(1) Run the keyutils testsuite against it.
(2) Using the AF_RXRPC and RxKAD modules to test keytype removal:
Load the rxrpc_s key type:
# insmod /tmp/af-rxrpc.ko
# insmod /tmp/rxkad.ko
Create a key (http://people.redhat.com/~dhowells/rxrpc/listen.c):
# /tmp/listen &
[1] 8173
Find the key:
# grep rxrpc_s /proc/keys
091086e1 I--Q-- 1 perm 39390000 0 0 rxrpc_s 52:2
Link it to a session keyring, preferably one with a higher serial number:
# keyctl link 0x20e36251 @s
Kill the process (the key should remain as it's linked to another place):
# fg
/tmp/listen
^C
Remove the key type:
rmmod rxkad
rmmod af-rxrpc
This can be made a more effective test by altering the following part of
the patch:
if (unlikely(gc_state & KEY_GC_REAPING_DEAD_2)) {
/* Make sure everyone revalidates their keys if we marked a
* bunch as being dead and make sure all keyring ex-payloads
* are destroyed.
*/
kdebug("dead sync");
synchronize_rcu();
To call synchronize_rcu() in GC phase 1 instead. That causes that the
keyring's old payload content to hang around longer until it's RCU
destroyed - which usually happens after GC phase 3 is complete. This
allows the destroy_dead_key branch to be tested.
Reported-by: Benjamin Coddington <bcodding@gmail.com>
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: James Morris <jmorris@namei.org>
0c061b5707